Applications of deep learning in the analysis of optical coherence tomography images for glaucoma-related diagnostics

Abstract

Glaucoma is an optic neuropathy and the leading cause of irreversible blindness worldwide. Imaging of the ganglion cell complex and retinal nerve fiber layer with optical coherence tomography (OCT) is a noninvasive, high-resolution means of diagnosing and quantitatively monitoring glaucoma. In the anterior segment, OCT can also be used to assess the anterior chamber angle and identify angle closure, a risk factor for glaucoma. The interpretation of OCT images for accurate diagnosis requires expert-level knowledge of both the technology and glaucoma. Deep learning (DL) is a subfield of artificial intelligence (AI), which is gaining prominence in health care for its ability to interpret images and approximate clinician judgment. This review summarizes recent research that demonstrates how DL can contribute to the analysis of OCT images in glaucoma. Deep neural networks can assist clinicians in checking the quality of OCT scans, quantifying the thickness of optic nerve tissues, evaluating the anterior chamber angle, diagnosing glaucoma, and detecting the progression of existing glaucoma. As further work expands on the generalizability, equity, and explainability of these DL techniques, AI-driven clinical support tools may become available for glaucoma diagnostics.

Introduction

Glaucoma is an optic neuropathy and the leading cause of irreversible blindness worldwide.[1,2] The symptomatic phase of the disease is marked by visual field loss. Introduced in 1991, optical coherence tomography (OCT) was a revolutionary technology with the ability to measure the retinal nerve fiber layer (RNFL) and detect pathologic changes in nerve structure that occur prior to irreversible visual field loss.[3,4,5,6] Since then, OCT has become an essential test in diagnosing many conditions pertaining to the retina, choroid, and optic nerve. Its applications have also extended to the anterior segment, where it can be used to image corneal, scleral, and conjunctival disorders, as well as assess the anterior chamber angle in the workup of glaucoma.

The contemporary world has seen artificial intelligence (AI) technology rapidly proliferate in the last decade. AI has enabled everyday computers to begin performing tasks requiring human-like intelligence. AI-based clinical tools have reached the market with approval by the U.S. Food and Drug Administration and European conformity (CE) marking, such as LumineticsCore (formerly IDx-DR), which screens fundus photographs for diabetic retinopathy.[7] At present, these technologies have largely been applied to fundus photographs for the detection of diabetic retinopathy, age-related macular degeneration, and glaucoma.

Intensive research is underway to design AI tools that assist in the evaluation of OCT for glaucoma diagnoses. These include various aspects of OCT assessment, from quality correction to interpretation, in anterior- and posterior-segment applications. In this review, we summarize the current state and future capabilities of AI in OCT evaluation as it pertains to glaucoma.

A Brief Background on Artificial Intelligence

AI is the automation of tasks that require human intelligence, such as visual perception, speech recognition, and decision-making. While this definition seems broad and subjective, it may be demonstrated more objectively by the Turing test, which states that if an observer interacting with an AI is unable to determine if it is a human or an artificial agent, the AI is intelligent. In the context of this article, an algorithm may be considered intelligent if it can make clinical decisions that are comparable to those of a trained glaucoma specialist.

Machine learning (ML) is a subdiscipline of AI that concerns itself with learning from data to make predictions. Conventionally, ML used statistical models like linear and logistic regression to fit models to data that have been prepared in tabular form by data scientists who have specifically selected and formatted the data elements or features included. Whereas these ML techniques are considered shallow learning for their ability to optimize parameter selection from curated data, deep learning (DL), a subfield of ML that is prominent in healthcare, goes a step further by incorporating both feature engineering and shallow learning. In other words, DL finds the optimal combination of features to model data and then directs these features to a final output layer to make a prediction. A major strength of DL is that it can handle nonlinear data flexibly through its process of feature combination.

DL has a burgeoning role in image classification, which makes it the key technology for this discussion of the applications of AI to OCT image analysis in glaucoma. The most widely applied DL architecture is the convolutional neural network (CNN). In processing an image like a fundus photograph, a neural network grossly passes its pixel values through a layer that extracts meaningful combinations of pixels (e.g., lines, contours, and colors) and passes these as features to other layers that combine those features into more complex features (e.g., a representation of an optic disc), before passing them to an output layer for classification [Figure 1]. Supervised learning allows an algorithm to tune its parameter weights for different features throughout the various layers to make predictions that better fit training data (i.e., a set of images labeled by their ground truth classifications). CNNs can then be tested for performance on independent testing data. An algorithm may be considered better suited for general use if it performs well on testing data that are independent from the training data, for instance, data collected from an external clinic or research center. There are several standard architectures of CNNs that are utilized by research included in this review, such as DenseNet, gNet3D, Inception, ResNet, and VGG; however, the specific details of these architectures are beyond the scope of this article. Several studies also employed ensemble learning, which averages the outputs from several CNNs to give a single prediction.

Figure 1.

A schematic representation of a convolutional neural network (CNN) for glaucoma classification. Note that a CNN often has multiple convolutional and pooling layers, which are simplified in this diagram

A key challenge to the implementation of DL-based tools in the clinical setting is the black-box problem. Simply put, a clinician cannot observe the neural layers to understand how a CNN made a prediction. Several important advancements have been made in this area to make DL more explainable. Gradient-weighted Class Activation Mapping (Grad-CAM) is a common tool used by studies included in this review. These maps are essentially heat maps that highlight features in images that were heavily weighted by the DL algorithm in defining a class (e.g., ”glaucoma” in a prediction of diagnosis). A quantitative method of saliency mapping has also been used, termed Testing with Concept Activation Vectors, which reports a quantitative measure of the weight placed on each feature.

How is Optical Coherence Tomography Used in Glaucoma?

OCT contributes to the evaluation of two ocular structures that are essential in the evaluation of glaucoma: ganglion cells and their axons and the anterior chamber angle. Measurement of the former is an important test for the diagnosis and longitudinal monitoring of the disease. The latter helps define the type of glaucoma and guide treatment.

Glaucomatous optic neuropathy is characterized by damage to the ganglion cell axon, atrophy of the axonal segment within the RNFL, and death of the ganglion cell of origin.[8] Axonal damage occurs preferentially at the superior and inferior poles of the optic nerve head, although the patterns of early damage can be diverse and eventually lead to diffuse thinning of the RNFL. In the contemporary assessment of glaucoma, OCT provides a quantitative measurement of the thickness of the ganglion cell complex as a complement to the visual examination of the optic nerve head and RNFL.[5,9]

Commercially available OCT machines generate segmentation of the RNFL and ganglion cell-inner plexiform layer (GCIPL) via unsupervised algorithms, such as thresholding, edge detection, active contours, and graph-based methods, allowing the automated measurement of the thickness of the ganglion cell complex and RNFL across the area of imaged retina.[10,11] RNFL and GCIPL thickness analyses then average sectoral and global thickness measurements and compare these averages to normative databases, reference databases of measurements from healthy control eyes. The averages are classified as “within normal limits” (within the 95^th percentile of thicknesses in the normative database), “borderline” (in the 95^th–99^th percentile), or “outside of normal limits” (above the 99^th percentile). Studies have demonstrated that these classifications are highly specific and sensitive for the diagnosis of glaucoma.[12,13]

Sequential thickness measurements over time can be tracked to assess disease progression, which can be differentiated from noise in repeat measurements and changes due to aging.[14,15,16] Newer tests using OCT consider deeper structures of the optic nerve, such as the Bruch’s membrane opening (BMO) minimum rim width and direct measurements of the lamina cribrosa.[17,18] Investigators have even suggested that OCT testing alone may be sufficient for diagnosing glaucoma.[19] Despite recent advances, many challenges remain in the interpretation of these results.

Quality Assurance and Segmentation in Optical Coherence Tomography

The use of OCT in assessing the optic nerve tissue for glaucomatous damage relies on the accurate delineation of the peripapillary RNFL and macular GCIPL. Despite improvements in imaging quality and segmentation algorithms, errors in segmentation remain common, occurring in 19.9%–46.3% of studies.[20] Errors on average lead to thinner measurements of RNFL thickness, which result in the overestimation of a patient’s likelihood of having glaucoma.[20] AI has been studied in several capacities to improve automated measurement of these tissues by OCT.

At the most basic level, Jammal et al. designed an AI tool to alert clinicians to the presence of a segmentation error in the automated OCT thickness report.[21] In their study, a CNN was trained on OCT B-scans with automated segmentations by the machine’s native software that had been labeled by human graders as containing an error or not. The DL algorithm achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.979 (95% confidence interval: 0.974–0.984), which improved when considering only “severe” segmentation errors that human graders determined to have a severe effect on RNFL thickness measurements. CAMs were built over the input images, giving human reviewers an idea of where the automated segmentation erred.

Investigators have tasked AI directly with performing segmentation. DL algorithms have been successful in segmenting multiple features of the optic nerve and peripapillary region, such as the RNFL, BMO, and lamina cribrosa. A complete review of the methods has been reported elsewhere and is beyond the scope of this article.[22] Overall, good results have been found. DL algorithms have also been uniquely capable of measuring the BMO and segmenting deeper tissues of the lamina cribrosa, where the signal-to-noise ratio is low in OCT. In terms of conventional RNFL measurement, several authors used DL to segment the RNFL and found thickness measurements to be equivalent with DL-derived and human manual segmentation.[23,24] In these studies, segmentations are output and can be manually reviewed by clinicians as a quality check.

A final approach to segmentation is to bypass it. Mariottoni et al. created a DL algorithm that used a segmentation-free process for measuring RNFL thickness.[25] A CNN was trained to predict average RNFL thickness from peripapillary OCT circle scans that were unsegmented. In testing high-quality OCTs, the DL algorithm performed at least as well as conventional segmentation algorithms; however, in tests of OCTs with segmentation errors or artifacts, the DL algorithm outperformed conventional segmentation when compared to automated measurements from good-quality OCTs from the same patients on the same day.

Machine to Machine

A number of studies have used AI to predict visual field results from OCT data. The visual field is the gold standard test for assessing the function of the optic nerve and its progressive degradation from glaucoma. The visual field is a time-consuming test to perform, which requires significant attention from patients and is susceptible to test-to-test variability and unreliable results. Predicting visual field results is of interest to bypass these limitations, when providing a familiar and interpretable result for clinicians. Investigators have used AI to predict both global outcomes of visual field testing, such as the presence of a glaucomatous visual field defect or the mean deviation of the test, as well as pointwise sensitivities, which represent the complete information obtained from a visual field test.

In terms of global measurements, Huang et al. demonstrated that CNNs outperform shallow ML methods in prediction.[26] Their CNN trained on peripapillary RNFL thickness measurements split into 64 averaged sectors and achieved good results in predicting visual field mean deviation in internal and external testing. Christopher et al. compared a DL algorithm trained on RNFL thickness maps, en face images from the Spectralis OCT, and confocal scanning laser ophthalmoscopy and had success in predicting the presence of a glaucomatous visual field defect.[27] Interestingly, they found that training on en face images from the Spectralis machine yielded the highest accuracy.

Several investigators have predicted pointwise visual field sensitivities. Xu et al. and Asaoka et al. used thickness maps of several macular layers to make pointwise predictions of 10-2 visual field sensitivities with good results and demonstrated that DL outperformed non-DL methods of prediction.[28,29] They additionally were able to train the DL algorithm to perform a second task of predicting a future visual field when passed a series of OCT thickness plots and visual fields. Other positive results have been published by Kamalipour et al., where CNN-based prediction of 10-2 sensitivities from RNFL thickness was reported, and by Park et al., where CNN-based prediction of 24-2 sensitivities from combined RNFL and GCIPL thickness maps was reported.[30,31] While it is difficult to compare results across studies due to differences in testing methodologies and study cohorts, investigators have consistently demonstrated that DL algorithms outperform conventional ML approaches.

While the studies discussed to this point have used segmented data, several investigators have used unsegmented data to avoid the limitations of automated segmentation discussed in the prior section. Mohammadzadeh et al. assessed the effect of training a CNN on raw OCT volumes of the macula, as opposed to GCIPL thickness maps, on predictions of 10-2 sensitivities.[32] They found that training on OCT volumes produced more accurate predictions. Similarly, Chen et al. demonstrated that a CNN trained on raw OCT peripapillary volumes outperformed a CNN trained on RNFL thickness maps in predicting 24-2 visual fields.[33] Interestingly, they also found that using OCT volumes improved accuracy when RNFL thickness reached its measurement floor, which suggests that there may be information recognized by the CNN in the OCT volumes that is difficult for clinicians to decipher from standard RNFL reports.

Predicting a Glaucoma Diagnosis

Beyond measuring OCT and predicting visual field results, many investigators have used AI to predict a glaucoma diagnosis from OCT results. Recent focus has been on DL methods as early publications demonstrated improvements in diagnostic accuracy with DL over ML and hybrid DL/ML models, especially when testing in external datasets.[34,35,36] A number of inputs have been used for predicting a glaucoma diagnosis, including OCT reports, B-scans, and volumes.

Muhammad et al. applied hybrid DL/ML and DL algorithms to OCT reports.[36,37,38] The “Hood report” used for training includes RNFL and GCC thickness and deviation maps, RNFL average thickness measurements by quadrant and clock hour, a circle-scan image of the segmented circumpapillary RNFL with a thickness plot, and an en face infrared image of the retina (obtained from the OCT machine).[19] The end-to-end DL models performed well when tested on data from the same source as the training dataset. The models were also tested on data from an external source and, although all models performed worse than when tested on internal data, the DL models demonstrated less performance attenuation than the hybrid DL/ML models. Saliency testing demonstrated that the RNFL deviation map and the GCC thickness and deviation maps were most important, which correlated with expert eye-tracking data during review of the same OCT reports.[36]

Recent studies using OCT B-scans have used unsegmented scans for training. Thompson et al. trained a DL algorithm on raw peripapillary circle scans.[39] The algorithm achieved a high AUC-ROC (0.96) and performed significantly better than global and sectoral RNFL thickness measurements in predicting a glaucoma diagnosis. Most of the difference came from the classification of less severe, especially preperimetric, glaucoma, where thickness thresholds are less sensitive. It is worth noting that this finding has not been consistent across studies in external testing; for instance, Ran et al. found that RNFL thickness performed comparably to a CNN in external testing despite internal testing results that favored the CNN.[40] Severity differences in the distributions of glaucoma seen in these studies may factor into these results.

A multicenter study developed three-dimensional (3D) CNNs for predicting glaucoma diagnosis from raw OCT volumetric cubes of the optic nerve and peripapillary region. Noury et al. trained and validated a 3D CNN on a diverse dataset from Stanford University.[18] They then tested on external datasets from Hong Kong, India, and Nepal with moderate-to-good performance (AUC-ROC 0.80–0.94). Variation in performance was suspected to be from protocol differences in labeling glaucoma and in dataset variance in terms of the distributions of glaucoma severity, more advanced disease being easier for AI to recognize. The collaborating investigators from Hong Kong, Ran, et al., trained a 3D CNN on their own OCT volumes and found similar performance in external testing on the same datasets (AUC 0.893–0.897).[41] In a separate publication, an overlapping group of investigators led by Ran et al. trained a similar model to detect both glaucoma and myopic features from OCT volumes and found that the addition of a second task did not reduce the performance of predicting glaucoma.[40] The single- and multitask DL algorithms trained in that study, however, did not outperform average RNFL thickness in external testing. It is also worth noting that several of these studies reported lower sensitivity of the DL algorithms at specificities held constant to human graders, though corresponding AUC-ROC values were not statistically different.[18,41] Several of these studies implemented Grad-CAMs and identified that the lamina cribrosa, in addition to the superficial retina, was often indicated as an area of algorithm attention.[18,35,41]

While many of the aforementioned studies used datasets of strictly healthy and glaucomatous OCTs, Russakoff et al. also included OCTs from eyes suspicious for glaucoma to develop a CNN for predicting referable glaucoma, which they defined as having manifest or preperimetric glaucoma or being at high risk for glaucoma with suspicious optic disc changes.[42] The CNN was trained on OCT volumes of the macula. Performance was similar to their previously described study (AUC: 0.78–0.95) despite the increased complexity of the task.

Overall, there is promise in using AI to assist the diagnosis of glaucoma based on OCTs, although there are potential limitations to the method. First, external testing, while vital to demonstrate model generalizability, is not consistently performed. Second, reports exist of standard RNFL metrics outperforming AI in external testing. Finally, it is not clear how algorithms will perform in real-world settings where there may be more of a gray zone in between glaucoma and nonglaucoma, and the distributions of glaucoma severity may be variable (often skewed toward mild).

Progression Analysis

Detecting progression is a very different task from diagnosis. In some ways, it may seem simpler as it is essentially a change analysis from a baseline state. In other ways, it is much more difficult as detecting progression requires analysis of longitudinal data. There is also no consensus definition of progression on OCT, and test–retest variability and aging complicate the separation of true progression from noise. Far fewer studies have been performed in this area.

Early publications on this topic used ML. Belghith et al. leveraged a statistical approach using a Bayesian hierarchical model to generate change maps for RNFL measurements over time (derived from OCT volumes), then applied an ML algorithm to classify the change maps as progressing or not.[43] They designed the ML algorithm as a one-class classifier, meaning that it was trained only on nonprogressive control data. The same group, published under Bowd et al., later implemented a DL algorithm to perform the classification task when presented a change map between baseline and follow-up RNFL segmentations.[44] Using an unsupervised deep-learning autoencoder and a similar one-class training approach, they were able to achieve a high sensitivity and specificity for detecting progression, outperforming standard RNFL parameters.

In a different approach, Mariottoni et al. trained a CNN to predict a likelihood of progression when given pointwise RNFL thickness measurements from two OCT studies.[45] The model, which was learned from longitudinal studies, achieved an AUC-ROC of 0.938 and, at matched specificities, outperformed traditional trend-based analysis (e.g., linear regression) using global and sectoral average RNFL thickness measurements. The model also output heatmaps to identify locations in the RNFL scan where change likely occurred.

How is Optical Coherence Tomography Used in Evaluating the Anterior Segment in Glaucoma Care?

While the primary application of OCT in glaucoma has been for imaging the posterior segment, OCT technology has also been adapted for imaging the anterior segment. Qualitative and quantitative assessments of the anterior chamber angle, iris, and lens play important roles in glaucoma care.

Glaucoma subtypes are commonly divided based on the anatomical status of the anterior chamber angle.[46] The anterior chamber angle houses the trabecular meshwork, the outflow structure responsible for maintaining pressure homeostasis by draining aqueous humor from the eye. In the spectrum of angle-closure disease, there is appositional or synechial closure of the anterior chamber angle that reduces drainage and increases intraocular pressure (IOP).[47] Angle closure is a significant risk factor for glaucoma.[48] Therefore, treatments such as laser peripheral iridotomy (LPI) and lens extraction are indicated when angle closure is extensive, resulting in the development of peripheral anterior synechiae (PAS) or elevated IOP.[48,49]

Gonioscopy is the current clinical standard to assess the anterior chamber angle to detect angle closure. However, gonioscopy is subjective, qualitative, and poorly reproducible even when performed by glaucoma specialists.[50] These limitations contribute to gonioscopy being underutilized despite the key role it plays in the glaucoma evaluation, leading to delayed detection of primary angle-closure glaucoma (PACG) and higher prevalence of PACG-related blindness.[51,52,53,54] Anterior segment OCT (AS-OCT) provides a convenient alternative to gonioscopy due to its noncontact, quantitative nature.[55] More recent work suggests disagreements between the two assessment methods may favor AS-OCT, at least for quantifying angle biometrics.[56,57] One limitation of AS-OCT is that the trabecular meshwork cannot be directly visualized. Therefore, the detection of angle closure and derivation of biometric measurements relies on manual identification of the scleral spur, which delineates the posterior extent of the trabecular meshwork. The scleral spur detection process can be time-consuming and expertise dependent, especially when many images are obtained per eye to capture the anatomical variations of the angle.[58,59,60,61] Anterior chamber biometrics have been linked to IOP, anatomical mechanisms and progression of primary angle-closure disease (PACD), and response to LPI.[62,63,64,65,66,67,68] Thus, AS-OCT could play an important role in detecting angle-closure and risk-stratifying patients for PACG.

Measuring the Anterior Chamber

Many useful anterior chamber biometrics are measured in relation to the scleral spur, such as angle opening distance, trabecular iris space area, and anterior lens vault. Labeling of the scleral spur is a complex manual task and the scleral spur is not always evident to human graders.[58] Xu et al. approached this problem by training a CNN on labeled AS-OCT images to predict the Euclidean coordinates of scleral spurs.[69] The model performed well, with mean absolute prediction errors comparable to the intragrader variability of an expert grader. Pham et al. used an ensemble of CNNs to perform scleral spur localization and segmentation of AS-OCT images by tissue type (i.e., iris, cornea and sclera, anterior chamber, and background), which facilitated the complete automation of the biometric analysis process.[70] Based on predicted scleral spur locations, for which machine–human agreement approximated human–human agreement, they found a strong correlation between DL- and human-derived measurements of anterior segment biometric parameters (ICCs >0.86 for all parameters). In a separate publication, the same group, published under Soh et al., refined the DL algorithm and demonstrated that the predictions were significantly less variable than those from the Zhongshan Angle Assessment Program, which is a commonly used image analysis software for AS-OCT images that does not use AI.[71] Heidelberg Engineering recently received CE Mark approval to use a CNN to identify the scleral spur, which has been incorporated in its latest OCT machine with anterior segment capability. This algorithm performed well when compared to expert graders on an independent test set, achieving algorithm-expert variability in scleral spur localization that compared favorably to expert–expert variability and producing biometrics with algorithm-expert agreement that was stronger than novice-expert agreement when evaluating narrow angles (intraclass correlation coefficients ranging 0.746–0.979 vs. 0.146–0.929), suggesting a performance advantage to the algorithm over the common clinical user.[59]

Detecting Angle Closure

As in the posterior segment, many have bypassed the step of quantitative measurement and directly tasked AI with diagnosing anatomical angle closure. Fu et al. trained a CNN to categorize angle closure based on AS-OCT images, which were labeled as closed if there was iris-trabecular contact (ITC), defined as contact between the anterior iris and inner cornea anterior to the scleral spur.[72] When comparing this DL model to an ML model using extracted features (several of the biometric parameters discussed in the last section), they found that the DL algorithm reached a higher AUC (0.96 vs. 0.90) with improved sensitivity and specificity for identifying ITC. Yang et al. achieved a similar performance for a CNN-detecting ITC based on labeled AS-OCT images (AUC: 0.963).[73] An additional study from Shan et al. trained CNNs in attempt to further classify ITC into the different stages of PACD.[74] While performance was strong in differentiating PACD eyes from controls (AUC: 0.95 in external test), a task which is essentially detecting ITC, it was substantially weaker in separating primary angle-closure suspects from PAC/PACG (AUC: 0.64).

Investigators have also used DL to predict gonioscopic findings from AS-OCT images. Xu et al. trained a CNN to predict the probability of each gonioscopic angle grade (based on the modified Schaffer grading system) in individual AS-OCT images from each quadrant of the anterior chamber.[75] They assessed the generalizability of the algorithm and found a consistent AUC (0.894–0.922) across independent test datasets that varied by clinical setting and patient demographics.[50] They also used misclassifications by the CNN to elucidate differences between AS-OCT and gonioscopic definitions of angle closure.[76] Porporato et al. developed a CNN to evaluate the 360° angle configuration of eyes (128°CT cross-sections per eye) and achieved an AUC of 0.85 when comparing its detection of angle closure to gonioscopy.[77] Addressing more detailed angle findings, Li et al. trained a 3D-CNN to differentiate between ITC and PAS, as had been evaluated with dynamic gonioscopy and found strong performance on an independent test dataset (AUC: 0.902).[78] Wanichwecharungruang et al. were able to classify a diagnosis of plateau iris configuration, a subtype of angle closure established with ultrasound biomicroscopy (UBM), from AS-OCT images using a CNN, whereas human manual interpretation of the AS-OCT images was in poor agreement with UBM (sensitivity 87.9% vs. 56.47%–77.78% and specificity 97.6% vs. 48.94%–64.29%).[79,80]

Given that the accumulation of sufficient training data and labeling is time-consuming, Zheng et al. explored the effect of creating synthetic data to enrich training.[81] They trained a generative adversarial network to create synthetic AS-OCT images based on real ones. Human graders found the synthetic images to be of comparable quality to the real ones, though the scleral spurs were less visible in those images. DL algorithms trained on real and synthetic datasets to detect angle closure had comparable performance on a testing dataset (AUC: 0.97 vs. 0.94).

Conclusion

In this review, we summarized recent research that demonstrates how AI can contribute to the interpretation of OCT imaging as it pertains to glaucoma. Deep neural networks can assist clinicians in checking the quality of OCT scans, quantifying the thickness of optic nerve tissues, evaluating the anterior chamber angle, diagnosing glaucoma, and detecting the progression of existing glaucoma. Despite these promising results, several barriers to DL algorithm implementation remain. First, AI algorithms must be generalizable for widespread use. External validation is time- and labor-intensive, especially since there is a shortage of high-quality, publicly available datasets for training and testing of glaucoma algorithms. Insufficient external validation may lead to variability in algorithm performance depending on local differences in features or demographics. A diverse central repository of ocular images, analogous to ImageNet, could be helpful in benchmarking algorithm performance and confirming generalizability. Second, the use of DL is resource-intensive, requiring computing power that exceeds that of the typical desktop computer. This high resource demand means investigators must clearly demonstrate that DL outperforms less resource-intensive methods, like traditional ML. It also raises concerns about which healthcare settings will be able to access AI and at what cost. Third, additional buy-in is needed from providers, patients, and regulatory agencies for the adoption of AI-based tools in health care.

Further research is underway to address each of these issues. Larger studies with external validations in various care settings, challenged by diverse patient groups and indeterminant pathologies, will continue to improve generalizability. New techniques like federated learning may also improve generalizability and data availability when maintaining data privacy. Beyond these barriers, there is a gap for future work on the translation of AI algorithms into clinical care for the purposes discussed in this paper. A shift to the study of utility, including not only prospective efficacy but also adoption, cost, and clinical outcomes, will be crucial to advance AI into the clinic.

Data availability statement

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.

Conflicts of interest

The authors declare that there are no conflicts of interests in this paper.

Funding Statement

This study was supported by grants (R01 EY035677 and K23 EY032985) from the National Eye Institute, National Institutes of Health, Bethesda, Maryland, and an unrestricted grant to the Department of Ophthalmology from Research to Prevent Blindness, New York, NY. The research reported in this publication was also supported by the National Eye Institute of the National Institutes of Health under Award Number P30EY029220 (Data Science and AI Core). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.